Weiqiang Wang

Ant Group (China) · China

Unclaimed · AI-structured profileORCID OpenAlex

Ant Group researcher advancing efficient LLM inference, data-efficient RLVR post-training, and multimodal GUI agents

papers here

reviewed by Evolve

first-author

corresponding

—

works · OpenAlex

h-index · OpenAlex

Bibliometric figures come from OpenAlex (2026-10-16) and cover the author's entire public record, not just the papers listed here. Reproducibility signal comes from Evolve's own reviews.

Research profile

Output—

This axis would measure cumulative publication volume, venue mix, and per-year productivity from the counts-by-year series (which spans 1993–2025), but verified output aggregates are not yet available for this OpenAlex record, whose long span suggests homonym contamination. No score is assigned.

Impact—

Although the raw snapshot lists an h-index of 25 and an i10-index of 41, validated citation aggregates are not yet available for this record. This axis would measure citation-based influence and field-normalized impact, and it is deliberately left unscored rather than estimated.

Focus0/100

The focus axis scores 0, reflecting the OpenAlex record's scattered topic distribution: 'Algebraic structures and combinatorial models' (28 works) and 'Handwritten Text Recognition Techniques' (28) sit alongside 'Advanced Neural Network Applications' (25), 'Advanced Algebra and Geometry' (24), and 'Hand Gesture Recognition Systems' (17). This dispersion likely signals homonym merging in the record rather than diffuse research; the 2026 catalog papers are tightly clustered on LLM systems.

Collaboration67/100

With a collaboration score of 67, the record shows 28 distinct coauthors against 7 works in the last three years — a broad but managed network. The frequent-coauthor list spans twelve names, including Shenzhi Yang, Guangcheng Zhu, Kai Tang, Zhengqing Zang, and Xing Zheng, alongside academic partners such as Bo Han, Bo An, Lei Feng, and Junbo Zhao, consistent with industry–academia co-publishing from Ant Group.

Activity61/100

Activity scores 61. The counts-by-year series shows sustained output with twin peaks of 11 works in 2018 and 2022, plus 6 works in 2015 and 2020, and 5 in 2021 and 2023; 2024 dipped to 2 before recovering to 5 in 2025. The trailing three years total 7 works, indicating steady rather than high-volume recent output.

Reproducibility55/100

Reproducibility averages 55 across the 3 Evolve-assessed papers, all rated 'medium'. FastE is training-free but uses backbone-tuned thresholds; the GUI-agent work reports a 6.9→58.6% conflict-success gain yet relies on synthetic conflicts; DE-Venus is well-engineered on verl but shows small margins without variance checks. Consistent medium ratings suggest usable artifacts still needing independent replication.

Research summary

Weiqiang Wang is a researcher at Ant Group (China) whose recent work centers on making large language models cheaper to run and safer to deploy: efficient LLM inference, data-efficient post-training with verifiable rewards (RLVR), and multimodal GUI agents. The Evolve catalog includes three of his 2026 arXiv papers — FastE on readout-triggered token compression for LLM embedding inference, a conflict-aware termination framework for GUI agents, and DE-Venus, a verl-based data-efficient RLVR framework — in each case as a middle-list author rather than corresponding author. His OpenAlex record (ORCID 0000-0002-0004-7154) shows an h-index of 25 with an i10-index of 41, 28 distinct coauthors, and 7 works over the last three years; the output and impact axes lack sufficient verified data for scoring. Activity is sustained, with recorded peaks of 11 works in both 2018 and 2022 and 5 works in 2025. The record's topic distribution is unusually scattered — algebraic structures, handwritten text recognition, and gesture recognition sit alongside neural network applications — so this profile should be read as name-resolved from public bibliographic data, with possible homonym contamination in the record's long tail. Frequent coauthors include Shenzhi Yang, Guangcheng Zhu, Kai Tang, Bo Han, Bo An, and Junbo Zhao.

Collaboration pattern

The collaboration pattern is industry-anchored and broad: 28 distinct coauthors against 7 works in the last three years (collaboration score 67). The recurring roster mixes Ant Group colleagues — Guangcheng Zhu, Kai Tang, Zhengqing Zang, Xing Zheng, Yingfan Ma, and Bowen Song — with academic partners including Shenzhi Yang, Bo Han, Bo An, Lei Feng, Haobo Wang, and Junbo Zhao, a structure typical of industrial labs co-publishing with university groups. All three catalog papers are middle-position, non-corresponding contributions.

Research keywords

LLM embeddingstoken compressiontraining-free inference accelerationefficient LLM inferenceRLVRdata-efficient post-trainingGRPOGUI agentsmultimodal LLMsconflict-aware terminationactivation steeringinference-time intervention

Highlights

Training-free token compression for LLM embedding inference
FastE triggers prefix compression via readout alignment, cutting depth-wise redundancy without retraining; Evolve finds the broad gains credible while noting thresholds are tuned per backbone (reproducibility: medium).
Teaching GUI agents when not to act
Conflict-aware termination combines activation steering with prompting, lifting conflict-scenario success from 6.9% to 58.6% at roughly a 2.6-point cost on feasible tasks; Evolve flags that the conflict cases are synthetic.
DE-Venus: data-efficient RLVR on verl
A unified verl-based framework for reinforcement learning with verifiable rewards, targeting data efficiency through weak supervision and noisy-label handling; Evolve credits real savings but notes small margins lack variance checks and key methods are self-cited.

Timeline

20156 works are recorded in this year, among the higher annual counts in the OpenAlex record before the 2018 peak.
2018The record reaches a first peak of 11 works in a single year in the counts-by-year series.
2022A second 11-work peak appears, matching the 2018 high in the record.
20255 works are recorded; the trailing three years total 7 works, supporting the activity score of 61.
2026Three arXiv papers enter the Evolve catalog: FastE for efficient LLM embedding inference, conflict-aware termination for GUI agents, and the DE-Venus data-efficient RLVR framework.

Papers in the Evolve catalog

Frequent co-authors

Shenzhi Yang×1 Guangcheng Zhu×1 Kai Tang×1 Zhengqing Zang×1 Xing Zheng×1 Haobo Wang×1 Yingfan Ma×1 Bowen Song×1 Bo Han×1 Bo An×1 Lei Feng×1 Junbo Zhao×1

About this profile

This profile is AI-structured by Evolve.Pub from public bibliographic data (OpenAlex) and Evolve's own paper reviews. It is not written or endorsed by the researcher unless the profile is claimed. Radar dimensions are computed from real bibliometrics with fixed formulas — raw figures are shown alongside; missing dimensions mean insufficient data, never zero. Errors? Claim the profile to correct or hide it.

Bibliometrics come from a single OpenAlex record (identity: openalex) and may merge same-name authors; the topic mix spanning algebra, handwritten text recognition, and gesture recognition suggests possible homonym contamination in the long tail.
Output and impact dimensions are unscored because verified publication and citation aggregates are unavailable in the current data snapshot.
All three catalog papers are 2026 arXiv preprints with middle author positions and no corresponding-author role; no peer-reviewed venue status is captured in this profile.
Reproducibility signals cover only the 3 papers reviewed by Evolve; other works in the record were not assessed.